350 Hz, 80 ms - University of Amsterdam

Modelling the perceptual development
of phonological contrasts with
OT and Gradual Learning Algorithm
Paola Escudero, University of Reading
[email protected]
Paul Boersma, University of Amsterdam
[email protected]
25th Penn Linguistics Colloquium
March 3, 2001
Sound contrasts
 Sound contrasts and acoustic information
 An example of a sound contrast
What are the acoustic differences between
the two?
Two acoustic cues to
“ship” vs. “sheep”
250
300
350
400
450
500
50
75
100
125
Duration (ms)
150
175
Elspeth’s production environment
(Scottish English)
250
300
350
sheep
feeling
lead
sneaker
400
450
ship
lid
Snicker filling
500
50
75
100
125
150
Duration (ms)
175
Liz’s production environment
(Southern English)
250
sneaker
sheep lead
Snicker
300
feeling
ship
350
lid
filling
400
450
500
50
75
100
125
Duration (ms)
150
175
Elspeth’s and Liz’s
average production environments
280
/i/
320
/I/
/i/
360
400
440
480
/I/
60
72
84
96
Duration (ms)
108
120
Do Elspeth and Liz perceive [350
Hz, 80 ms] as “ship” or as “sheep”?
280
/i/
320
/I/
360
Liz
?
/i/
Elspeth
400
440
480
/I/
60
72
84
96
Duration (ms)
108
120
Why using the nearest production
prototype in perception?
 Answer: “likelihood
maximisation”:
choose the most likely produced category,
given a certain F1 & duration
“minimise the
probability of perceptual confusion”
 Functional principle:
How Elspeth and Liz perceive the
segments reliably
[350 Hz, 350 Hz 80 ms 80 ms 350 Hz
80 ms] not /I/ not /i/ not /I/ not /i/
/I/
*!
*
*
*

/i/
[350 Hz, 350 Hz 80 ms 80 ms 350 Hz
80 ms] not /i/ not /i/ not /I/ not /I/

*
/I/
/i/
*!
*
*
First stage of babies Elspeth and Liz
Baby Elspeth or baby Liz
Spectral reliance 0.5%, duration reliance -0.3%
280
320
360
400
440
480
60
72
84
96 108
Duration (ms)
120
How baby Elspeth learns
 Little Elspeth makes a mistake
when categorising [350 Hz, 80 ms]
[350 Hz, 350 Hz 80 ms
80 ms] not /i/ not /i/

/I/

/i/
*!
*
80 ms
not /I/
350 Hz
not /I/
*
*
How little Elspeth learns to perceive
“ship” and “sheep” reliably (1)
Elspeth after one month
Spectral reliance 80.9%, duration reliance 12.2%
280
320
360
400
440
480
60
72
84
96
108
Duration (ms)
120
How little Elspeth learns to perceive
“ship” and “sheep” reliably (2)
Elspeth after 2 months
Spectral reliance 96.9%, duration reliance 9.3%
280
320
360
400
440
480
60
72
84
96
108
Duration (ms)
120
How little Elspeth learns to perceive
“ship” and “sheep” reliably (3)
Elspeth after 4 months
Spectral reliance 96.4%, duration reliance 0.4%
280
320
360
400
440
480
60
72
84
96
108
Duration (ms)
120
How little Elspeth learns to perceive
“ship” and “sheep” reliably (4)
Elspeth after 10 months
Spectral reliance 98.6%, duration reliance 15.7%
280
320
360
400
440
480
60
72
84
96
108
Duration (ms)
120
How little Elspeth learns to perceive
“ship” and “sheep” reliably (5)
Elspeth after 100 months
Spectral reliance 97.8%, duration reliance 6.3%
280
320
360
400
440
480
60
72
84
96
108
Duration (ms)
120
How little Elspeth learns to perceive
“ship” and “sheep” reliably (6)
Elspeth after 1000 months
Spectral reliance 97.1%, duration reliance 6.5%
280
320
360
400
440
480
60
72
84
96
108
Duration (ms)
120
How little Liz learns to perceive
“ship” and “sheep” reliably (1)
Liz after one month
Spectral reliance 34.6%, duration reliance 39.7%
280
320
360
400
440
480
60
72
84
96
108
Duration (ms)
120
How little Liz learns to perceive
“ship” and “sheep” reliably (2)
Liz after 2 months
Spectral reliance 37.1%, duration reliance 55.2%
280
320
360
400
440
480
60
72
84
96
108
Duration (ms)
120
How little Liz learns to perceive
“ship” and “sheep” reliably (3)
Liz after 4 months
Spectral reliance 53.5%, duration reliance 52.7%
280
320
360
400
440
480
60
72
84
96
108
Duration (ms)
120
How little Liz learns to perceive
“ship” and “sheep” reliably (4)
Liz after 10 months
Spectral reliance 46.3%, duration reliance 85.0%
280
320
360
400
440
480
60
72
84
96
108
Duration (ms)
120
How little Liz learns to perceive
“ship” and “sheep” reliably (5)
Liz after 100 months
Spectral reliance 51.9%, duration reliance 56.6%
280
320
360
400
440
480
60
72
84
96
108
Duration (ms)
120
How little Liz learns to perceive
“ship” and “sheep” reliably (6)
Liz after 1000 months
Spectral reliance 54.0%, duration reliance 51.4%
280
320
360
400
440
480
60
72
84
96
108
Duration (ms)
120
The perception of real adult Elspeth
and Liz
ScottishAverage
[i]
[i:]
4.49
SouthernAverage
[i]
[i:]
5.2
4.09
3.71
[I]
[I:]
reliance 11%, weight 10%
[I]
[I:]
reliance 32%, weight 30%
What about L1-Spanish Isabel,
who moves to Scotland and then
to Southern England?
Isabel’s production environment
(Spanish)
280
/i/
320
360
400
440
/e/
480
60
72
84
96
Duration (ms)
108
120
Isabel’s adult perception (Spanish)
Isabel after 20 years in Peru
Spectral reliance 99.9%, duration reliance 2.2%
280
320
360
400
440
480
60
72
84
96
108
Duration (ms)
120
Isabel’s new production environment
(Scottish English)
280
320
/i/
I FA
360
400
440
480
/I/
60
72
84
96
Duration (ms)
108
120
How Isabel learns to perceive
“ship” and “sheep”
Isabel after 10 years in Edinburgh
Spectral reliance 98.8%, duration reliance 4.4%
280
320
360
400
440
480
60
72
84
96
108
Duration (ms)
120
Isabel’s new production environment
(Southern English)
280
/i/
320
/I/
360
400
440
480
60
72
84
96
Duration (ms)
108
120
How Isabel learns to perceive
“ship” and “sheep”
Isabel after 10 years in London
Spectral reliance 69.3%, duration reliance 44.5%
280
320
360
400
440
480
60
72
84
96
108
Duration (ms)
120
Three types of real L2 categorisation
mt
[i]
[i:]
4.4
4.8
[I]
[I:]
[i]
abg
3.5
[i:]
[i]
af
3.29
[i:]
5.29
reliance -10%
[I]
[I:]
reliance 40%, weight 46%
[I]
[I:]
3.4
reliance 87%, weight 88%
Spectral reliance only
Cue integration
Duration reliance only
mf
jg
[i]
[i:]
[i] 1.92
6.5
[i:]
[i]
ef
4.75
[i:]
4.67
4.38
2.5
[I]
[I:]
reliance -6%
[I]
[I:]
reliance 37%, weight 33%
[I]
[I:]
4.25
reliance 86%
Summary

For our perception model, we assumed:
1) Perception is handled by an OT grammar
2) Its acquisition is handled by the GLA
3) L2 learners start by copying their L1 grammar

With these assumptions, we can model:
1) L1 Scottish and Southern English
2) L2 Scottish and Southern English, partially
Conclusion
 Cue reliance depends on cue reliability, or
a) Differences in the production environment
account for differences in perception.
b) Changes in the production environment
lead to changes in perception.
 The functional principle underlying this
production-perception dependence is
“minimisation of perceptual confusion”.
 This functional principle follows from our
formal modelling of the perception grammar.
Thank you for your attention!